1. Ameet Soni* and Jude Shavlik Dept. of Computer Sciences Dept. of Biostatistics and Medical Informatics Craig Bingman Dept. of Biochemistry Center for Eukaryotic Structural Genomics Presented at the ACM International Conference on Bioinformatics and Computational Biology 2010 Guiding Belief Propagation using Domain Knowledge for Protein-Structure Determination

2. 2 Protein Structure Determination  Proteins essential to most cellular function  Structural support  Catalysis/enzymatic activity  Cell signaling  Protein structures determine function  X-ray crystallography is main technique for determining structures 2

3. 3 X-ray Crystallography: Background Electron-Density Map (3D Image) Interpret Protein Crystal X-ray Beam Protein Structure 3 Diffraction pattern FFT Collect

4. 4 Task Overview  Given:  A protein sequence  Electron-density map (EDM) of protein  Do:  Automatically produce a protein structure (or trace ) that is All atom Physically feasible 4 SAVRVGLAIM...

5. 5 Challenges & Related Work 1 Å2 Å3 Å4 Å Our Method: ACMI 5 ARP/wARP TEXTAL & RESOLVE

6. 6 ACMI Overview 6 - Background - Inference in ACMI-BP - Guiding Belief Propagation - Experiments & Results

7. 7 Our Technique: ACMI Perform Local Match Apply Global Constraints Sample Structure ACMI-SHACMI-BPACMI-PF 7 p k +1 ( b ) k+ 1 *1 p k +1 ( b ) k+ 1 *2 p k +1 ( b ) k+ 1 *M … b k b k-1 b k+1 *1…M a priori probability of each AA’s location marginal probability of each AA’s location all-atom protein structures

8. 8 Previous Work [DiMaio et al, 2007] 8

9. 9 ACMI Framework Perform Local Match Apply Global Constraints Sample Structure ACMI-SHACMI-BPACMI-PF 9 p k +1 ( b ) k+ 1 *1 p k +1 ( b ) k+ 1 *2 p k +1 ( b ) k+ 1 *M … b k b k-1 b k+1 *1…M a priori probability of each AA’s location marginal probability of each AA’s location all-atom protein structures

10. 10 Inference in ACMI-BP 10 - Background - ACMI Overview - Guiding Belief Propagation - Experiments & Results

11. 11 ACMI-BP 11  ACMI models the probability of all possible traces using a pairwise Markov Random Field (MRF) LEUSERGLYLYSALA

12. 12 ACMI-BP: Pairwise Markov Field LEUSERGLYLYSALA  Model ties adjacency constraints, occupancy constraints, and Phase 1 priors 12

13. 13 Approximate Inference  P(U|M) intractable to calculate, maximize exactly  ACMI-BP uses Loopy Belief Propagation (BP)  Local, message-passing scheme  Distributes evidence between nodes  Approximates marginal probabilities if graph has cycles 13

14. 14 ACMI-BP: Loopy Belief Propagation LYS 31 LEU 32 m LYS31→LEU32 p LEU32 p LYS31 14

15. 15 ACMI-BP: Loopy Belief Propagation LYS 31 LEU 32 m LEU32→LEU31 p LEU32 p LYS31 15

16. 16 Guiding Belief Propagation 16 - Background - ACMI Overview - Inference in ACMI-BP - Experiments & Results

17.  Best case: wasted resources  Worst case: poor information given more influence Message Scheduling 17 SERLYSALA  Key design choice: message-passing schedule  When BP is approximate, ordering affects solution [Elidan et al, 2006]  ACMI-BP uses a naïve, round-robin schedule

18. 18 Using Domain Knowledge 18  Idea: use expert to assign importance of messages  Biochemist insight: well-structured regions of protein correlate with strong features in density map  eg, helices/strands have stable conformations  Protein disorder - regions of a structure that are unstable/hard to define  ACMI-BP can use disorder to decide importance  Accurate predictors exist based on sequence alone

19. 19 Guided ACMI-BP 19

20. 20 Related Work  Assumption: messages with largest change in value are more useful  Residual Belief Propagation [Elidan et al, UAI 2006]  Calculates residual factor for each node  Each iteration, highest residual node passes messages  General BP technique 20

21. 21 Experiments & Results 21 - Background - ACMI Overview - Inference in ACMI-BP - Guiding Belief Propagation

22. 22 Message Schedulers Tested 22  Our previous technique: naive, round robin (BP)  Our proposed technique: Guidance using disorder prediction (DOBP)  Disorder prediction using DisEMBL [Linding et al, 2003]  Prioritize residues with high stability (ie, low disorder)  Residual factor (RBP) [Elidan et al, 2006]

23. 23 Experimental Methodology  Run whole ACMI pipeline  Phase 1: Local amino-acid finder (prior probabilities)  Phase 2: Either BP, DOBP, or RBP  Phase 3: Sample all-atom structures from Phase 2 results  Test set of 10 poor-resolution electron-density maps  From UW Center for Eukaryotic Structural Genomics  Deemed the most difficult of a large set of proteins 23

24. 24 ACMI-BP Marginal Accuracy 24

25. 25 ACMI-BP Marginal Accuracy 25

26. 26 ACMI-BP Marginal Accuracy 26

27. 27 Protein Structure Results 27  Do these better marginals produce more accurate protein structures?  RBP fails to produce structures in ACMI-PF  Marginals are high in entropy (28.48 vs 5.31)  Insufficient sampling of correct locations

28. 28 Conclusions  Our contribution: framework for utilizing domain knowledge in BP message scheduling  General technique for belief propagation  Alternative to information-based techniques  Our technique improves inference in ACMI  Disorder prediction used in our framework  Residual-based technique fails  Future directions 28

29. 29  Phillips Laboratory at UW - Madison  UW Center for Eukaryotic Structural Genomics (CESG)  NLM R01-LM008796  NLM Training Grant T15-LM007359  NIH Protein Structure Initiative Grant GM074901 Thank you! Acknowledgements 29

30. 30 Protein Structures: Background  Building blocks are amino acids (AKA residues )  Chain of amino acids form the primary sequence Alpha Carbon Sidechain Backbone 30

31. 31 Related Work  ARP/wARP [Morris, Perrakis, and Lamzin, 1999]  TEXTAL [Ioerger and Sacchettini, 2003]  RESOLVE [Terwilliger, 2003]  BUCCANEER [Cowtan, 2006] 31

32. 32 0.25 0.35 0.45 0.55 0.65 0.250.350.450.550.65 0.25 0.35 0.45 0.55 0.65 0.250.350.450.550.65 0.25 0.35 0.45 0.55 0.65 0.250.350.450.550.65 Previous Work ACMI-PF R free ARP/wARP R free Resolve R free Textal R free 32

33. 33 ACMI-SH: Templates …SAW C VKFEKPADKNGKTE… Protein DB 33

34. 34 ACMI-SH: Fast Rotation Search pentapeptide fragment from PDB (the “template”) electron density map calculated (expected) density in 5A sphere map-region sampled in spherical shells template-density sampled in spherical shells sampled region of density in 5A sphere 34

35. 35 ACMI-SH: Fast Rotation Search map-region sampled in spherical shells template-density sampled in spherical shells template spherical- harmonic coefficients map-region spherical- harmonic coefficients correlation coefficient as function of rotation fast-rotation function (Navaza 2006, Risbo 1996) 35

36. 36 Backbone Model Potential Constraints between adjacent amino acids: =x 36

37. 37 Constraints between nonadjacent amino acids: Backbone Model Potential 37

38. 38 Observational (“amino-acid-finder”) probabilities Backbone Model Potential 38

39. 39 Publications 39  F. DiMaio, A. Soni, and J. Shavlik, “Machine learning in structural biology: Interpreting 3D protein images,” in Introduction to Machine Learning and Bioinformatics, 2008.  F. DiMaio, A. Soni, G. N. Phillips, and J. Shavlik, “Improved methods for template matchingin electron-density maps using spherical harmonics,” in Proceedings of the 2007 IEEE International Conference on Bioinformatics and Biomedicine  F. DiMaio, A. Soni, G. N. Phillips, and J. Shavlik, “Spherical-harmonic decomposition for molecular recognition in electron-density maps,” International Journal of Data Mining and Bioinformatics, 2009.  F. DiMaio, D. Kondrashov, E. Bitto, A. Soni, C. Bingman, G. Phillips, and J. Shavlik, “Creating protein models from electron-density maps using particle-filtering methods,” Bioinformatics, 2007.  E. S. Burgie, C. A. Bingman, S. L. Grundhoefer, A. Soni, and G. N. Phillips, Jr., “Structural characterization of Uch37 reveals the basis of its auto-inhibitory mechanism.” In preparation, PDB ID: 3IHR.

40. 40 Testset Density Maps (raw data) Density-map resolution (Å) Density-map mean phase error (deg.) 15 30 45 60 75 1.02.03.04.0 40

41. 41 ACMI-PF Overview  Particle refers to one layout of protein subsequence  An iteration alternately places, for each of N particles  C α position b k +1 given b k  All sidechain atoms s k given b k -1: k +1 bkbk b k+1 sksk b k-1 41

42. 42 ACMI-PF: Backbone Step (1) Sample M b k +1 ’s from empirical C  - C  - C  pseudoangle distribution b k b k+1 *1…M b k-1  place b k i +1 given b k i, b k i -1 (for particle i ) 42